Three-Dimensional Multi-Task Deep Learning Model to Detect Glaucomatous Optic Neuropathy and Myopic Features From...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucomatous optic neuropathy (GON) and myopic features (MF) are significant eye conditions.
• Spectral-domain optical coherence tomography (SDOCT) is crucial for diagnosing these conditions.
• Accurate and simultaneous detection of both GON and MF remains a challenge.

Purpose of the Study:

• To develop and evaluate a multi-task three-dimensional (3D) deep learning (DL) model for simultaneous detection of GON and MF.
• To assess the model's performance using SDOCT volumetric scans.
• To compare the DL model against traditional metrics like average RNFL thickness.

Main Methods:

• A ResNet-based 3D DL model was developed to output GON and MF presence (Yes/No).
• The model was trained and validated on SDOCT scans from a tertiary eye hospital.
• External validation was performed on five independent datasets from different international eye centers.

Main Results:

• The 3D DL model achieved superior internal validation performance for GON detection compared to average RNFL thickness (AUROC 0.949 vs. 0.913).
• External testing showed comparable performance between the DL model and RNFL thickness for GON detection.
• The DL model demonstrated generalizable performance for MF detection (AUROC 0.855-0.896) across all datasets.

Conclusions:

• The multi-task 3D DL model exhibits high generalizability for detecting both GON and MF.
• The presence of myopic features did not generally impact the accuracy of glaucomatous optic neuropathy detection by the model.